Mask defect inspection tool duel: KLAC vs. AMAT
06/01/2008
As chip feature sizes have fallen to less than one–third of the exposure wavelength, the advanced photomask has become more and more of an optical element. Mask patterns now contain sub resolution assist features (SRAFS) to enhance printing???hopefully without appearing in the final image???and new pixelated and inversion mask geometries bear little resemblance to the design intent (Figure). In addition, the mask error enhancement factor (MEEF, now >3 for some levels) implies that tiny edge placement errors or defects on the mask may cause disproportionate distortions of the image. Chip manufacturers might hope that careful fabrication and inspection would eliminate all printable defects, but it is hard to tell what is printable in a given process. Meanwhile, a new class of crystalline defects (which condense from the air) has appeared. Since one defect on one mask can ruin an entire production run, maskmakers and users have become justifiably paranoid. In such an environment, who would want to be responsible for mask inspection tools?
 |
As it happens, both KLA–Tencor and Applied Materials introduced new mask defect inspection capabilities at Photomask Japan in April: the Applied Aera2 aerial image inspection system and the Wafer Plane Inspection (WPI) detector for the K–T TeraScanHR. Both propose to solve the problem of identifying printable defects and rejecting numerous harmless anomalies, while speeding inspection, but they operate on very different principles.
TeraScanHR with WPI
The KLA–Tencor TeraScanHR uses a large numerical aperture microscope to capture high–resolution transmission and reflection images of the photomask (with pixel sizes down to 72nm at 4?? mask scale) that contain far more detail than can be printed on the wafer. The WPI software then takes this information and applies computational lithography methods to simulate both the aerial image projected by an exposure tool and the exposure pattern in resist. By comparing the aerial and resist images of different die on a reticle (die–to–die mode), printable defects can be separated from the nonprintable, according to a paper presented by Rajesh Nagpal and a team from Intel Mask Operations and KLA–Tencor. Because the final criteria are based on lower–resolution stepper–simulation images, 90nm pixel size on the WPI–enabled TeraScan captures all relevant defects???and rejects most false defects, speeding inspection by 40% compared to using the 72nm pixel size. Review of suspicious locations found by WPI on a Zeiss AIMS optical aerial image review tool (considered the gold standard of photomask inspection) revealed a good correlation???22 of 25 located were indeed printable, according to the authors.
Both line–and–space and contact–hole masks were inspected, for a total of 12 advanced node reticles. Hard defects (which always print), soft defects (which only appear at the edges of the focus exposure window), and reticle haze (from airborne contamination) were identified. The longest reported inspection time with 90nm pixels was 89 minutes. The availability of high–resolution reflected and transmitted images facilitated the evaluation of low contrast anomalies, such as haze, according to the paper. Left to itself, the high–resolution system with its 90nm pixel would have reported 319 potential defects, whereas the WPI system located 145, and reported 50 as yield–affecting.
WPI leverages innovations in computational lithography (as used for OPC verification by Brion, etc.) to extract information from the high–resolution reticle images captured by the familiar TeraScan inspection system. But first it must reconstruct the actual photomask structure, which is not trivial since phase–related information does not automatically appear in the transmitted or reflected images. Both phase–shifting mask structures and semi–transparent crystalline affect phase in yield–impacting ways. The exposure tool model (incorporating off–axis illumination and other key parameters) propagates the mask pattern to the wafer plane, forming an aerial image. Then a resist model (currently a simple threshold) converts the aerial image intensity into a predicted pattern of circuit features which can be compared to the desired structure. “Defects” are located where the critical dimensions of these predicted wafer patterns differ from those desired by some specified amount. While die–to–die mode has been demonstrated for the WPI detector, computational methods should also be able to simulate aerial images directly from the mask–design database. Defects identified on WPI can be readily correlated with high–resolution TeraScanHR images to facilitate identification and process improvement.
Aera2
The basic problem with high–resolution mask inspection, according to Mark Wagner, head of Applied Materials’ mask inspection division, is that the signal produced by a defect does not correlate with the defect’s impact on the wafer CD (Fig. 2). However, the distortion of the aerial image projected by a stepper emulator does correlate with the printed resist variation. This has long been known, and is the basis for the operation of the Zeiss AIMS mask review tool. The challenge is to capture data quickly enough and with sufficient fidelity to inspect an entire 6–in. photomask under likely focus conditions.
The Aera2 captures the light diffracted from a mask as would an exposure tool, but with lenses that magnify the image on a CCD type detector rather than de–magnifying it onto photoresist (Fig. 3). The illumination conditions can be set to match those of a stepper and the Aera2 numerical aperture is one–fourth that of the exposure tool, far less than that of a high–resolution mask inspection system. The exposure wavelength (193nm these days) must be correct, and inspecting the entire mask requires scanning and processing the images quite rapidly. That means using an excimer laser operating at kHz pulse rates, with limited pulse averaging, just like a scanner. Small mask defects cause subtle localized changes in the brightness of the aerial images, which are captured and digitized. The Aera2 thus compares an optically magnified version of the aerial image at the wafer with what would be needed for proper yield (by applying a threshold resist model), whereas the WPI software compares digitally computed stepper simulations.
 Figure 3. Optical emulation of lithography: pupil shape, NA, and σ. |
Wagner told SST that the Aera2 separates the printing defects from the nuisance and false defects directly, avoiding simulation and review steps. He claimed that the Aera2 shows twice the throughput of high–res inspection technology, resulting in the shortest time to decision.
The Aera2 also can perform a survey to capture the mask contribution to CD uniformity by measuring the brightness of every feature on the mask???5??108 measurements done in parallel with the conventional mask inspection. Subtle variations in mask processing are known to cause feature dimensions to vary by a few critical nanometers across the reticle. The IntenCD measurements on the Aera2 reveal ~1% nonuniformities on minimum CD features in regions as small as 10μm ?? 10μm, according to Wagner, facilitating improved mask process control (and possible downstream correction using technology from Pixer). He noted that the first shipment of the Aera2 was one year ago and reported that there are now several in the field, one replacing an older version at a major MPU manufacturer.
Since the aerial images now have simpler geometry than some OPC mask designs, the time may have come for aerial image inspection. Or it may be that simulation of the aerial image based on high–resolution mask inspection may prove sufficient. Pixelated masks, contact reticles with intersecting assist slots for random logic, and “loopy” inversion designs have all proved challenging for traditional systems to inspect, leading to desensitized defect inspection criteria or regions labeled “DNI” for “do not inspect.” One result has been uncertainty in chip yield as uncorrected mask anomalies become “killer” defects, but others have been over–priced or over–specified reticles, restrictions on OPC methodologies and outright fear.
If the Aera2 or WPI performs as advertised, it may lower reticle costs, help make chip yield more predictable at 45nm and 32nm, and enable even more effective RET strategies. If so, both companies will have done the semiconductor industry a service. On the other hand, the tight economics of the maskmaking and maskmaking–equipment segments may not reward their contributions or sustain continuing innovation in optical reticle manufacturing. ???M.D.L.
Toshiba, Ponte Solutions team up on full–chip VCE modeling
Earlier this year, Michael Buehler–Garcia, Ponte Solutions’ vice president of marketing and business development, discussed its strategy of taking DFM to the IP level, and indicated the company was completing a physics–based etch model for vias/contacts and poly/metal. The need for such a model is a result of the impact of pattern density variability on etch–induced pattern transfer, which requires that the process be accurately modeled and corrected at 45nm and beyond.
At the time the story was written, Ponte noted that its etch solution had been in beta evaluation with a major Japanese IDM during Q4 2007, with evaluations to continue into Q1 2008. Now, with the release of news about the joint effort on March 31, the major Japanese IDM had a name: Toshiba, specifically, Toshiba’s Corporate Manufacturing Engineering Center (TCMC). The companies are working together to develop a via/contact etch (VCE) model.
The significance of the die–level etch process modeling, as Ponte’s chief scientist, Valeriy Sukharev, told SST, is that there has been a missing link between reactor (or wafer) scale and feature scale simulations of the etch process. Because there is a difference of 6–7 orders–of–magnitude between the wafer size and a layout feature size, a die–level model is necessary, as it provides a link between wafer–level and feature–level simulation tools, and a way to model layout–induced intra–die etch variations. Such a model is usable during the design and MDP (mask data prep) stage to reduce the impact of design/pattern density induced etch variability.
The full–chip VCE model addresses fluxes of neutral radicals, etch rates, CD variations, and etch hotspots. “Once the latter are calculated for the full chip, our VCE can detect and report etch hotspots based on the fab defined thresholds of acceptable variations,” said Sukharev. He added that previous efforts to address non–uniform pattern density effects in etch processing at the design and MDP stages were rule–based. “But the more advanced the geometry, the more rules are needed. People tried to introduce proximity to first, second, and third neighbors, etc., and the amount of information and complexity kept growing.”
Sukharev noted that Ponte avoided the pitfalls of a rules–based approach by using across–die radical distribution considerations. “Combining these with the aspect–ratio–dependent intra–feature radical transport resistance allows us to predict the layout shapes’ distortion caused by etch–assisted pattern transfer,” he said. “We do not need to introduce all the proximity factors separately.” He noted that all the information about the die layout is implicit in the solution, so there is no need for the analyzed etch step to be run on a specially designed test chip.
Ara Markosian, Ponte’s CTO, told SST that the rule–based approach can also be used (and is currently used) as a “full–chip die–level” approach. “The point is that the rule–based solutions become insufficient for the 45nm node and below,” he said. “Only a few measurements are necessary to calibrate/tune the model to a process, which may translate to different savings depending on a fab’s turnaround time.” ???D.V.
AMAT’s cleaning cluster
Making photomasks has historically been incredibly difficult yet profitless, and maskmaking equipment has to be innovative to handle the creation of phase–shift masks with sub–resolution assist features (SRAF). Cleaning has always been essential for mask manufacturing, but there is even more concern with mask cleanliness today. Fabs (particularly memory) may soon start cleaning masks in–house during production use to prevent haze buildup. Also, both EUV and NIL could be in volume production within five years, and since neither approach allows for the use of a pellicle, the mask would almost certainly need to be cleaned in production.
 Figure 4. Schematic of Tetra cluster with dry/wet chambers. |
Addressing this potential need for mask cleaning is Applied Materials with its new Tetra Reticle Clean tool, based on a cluster of dry and wet processing chambers around an atmospheric robot handler mounted on a linear track. A top–down schematic provided by the company (Fig. 4) shows a typical configuration with three wet and one dry processing chambers. The wet chambers have in–situ drying capability such that masks are handled “dry–in/dry–out” by the robot.
The wet chamber cleans both sides of the mask simultaneously. The mask is held at corners to float it over a chuck, which supplies cleaning chemistry as well as megasonic energy. The megasonic energy is coupled through the backside aqueous chemistry to the mask, so that uniform cavitation is supplied to the topside through the mask itself. Topside cleaning also occurs due to micro–droplets with kinetic energy from a point–source rastering across the spinning topside. SC1 plus surfactants, or ozone/DIW chemistries can be used.
The company has shown 99% particle removal efficiency (PRE) on blank masks, and zero damage seen on patterned masks with sub–50nm SRAFs, which would be used for 32nm node processing. Since sulfur is known to contribute to haze growth, all chemistries, including those used for resist–strip, are sulfur–free.
The dry (or plasma) etch chamber uses an inductively coupled plasma (ICP) source to provide remote RF energy, as well as a gas distribution plate through which ions diffuse to reach the heated processing pedestal.
The company claims nine current customer engagements–which corresponds to every “Tier–1” mask shop in the world. Currently running in production at 45nm, the tool is reportedly cutting 30%–50% of the haze problem at a commercial fab. ???E.K.
Thermo–PV: Heat makes photons, too
The vast majority of photovoltaic (PV) installations capture photons emitted by the sun, but many other sources of electro–magnetic radiation exist. Combustion of any kind emits photons in the infrared portion of the spectrum???so waste heat produced by diesel generators, internal combustion engines, and industrial processes offers a potential source of radiation for PV cells with the appropriate bandgap.
Some systems generate electricity as a by–product of combustion, generally by mechanical means. Industrial waste heat powers steam turbines. In engines, some energy from the drive train may be used to drive an alternator, in which the rotation of a magnet generates electricity for use by the vehicle.
In both these cases, the conversion from thermal to mechanical to electrical energy increases system complexity and introduces mechanical and other losses. Such systems are also difficult to miniaturize for portable applications. While microturbines can generate a high power density, they require precise, high speed, moving parts, notes W.M. Yang of the U. of Singapore. Creating and assembling millimeter–scale microturbines is quite challenging [1]. Moreover, such mechanical systems make noise and are subject to friction and wear.
Military applications would benefit from a lightweight, near–silent source of electricity???e.g., unmanned reconnaissance vehicles that operate cameras and sensors, with heavy limited–charge batteries that constrain the vehicle’s range.
Conventional diesel generators also convert thermal energy to electricity by way of a mechanical system, and are often used in parallel with heat sources for cooking and climate control???a way to generate electricity as a by–product of heating might offer substantial weight savings.
Applications like these are driving interest in thermophotovoltaic (TPV) devices, which place photovoltaic cells in close proximity to a combustion heat source (usually 1000???1600K, or 1340?????2420??F). They have no moving parts, substantially reducing weight and complexity (and size) compared to conventional generators. Yang’s group, for instance, demonstrated a 0.92W unit with a 3.0mm–diameter micro–combustor. Such small units can deliver more output power/unit volume than larger units because of their high surface–to–volume ratio.
The most significant difference between TPV and conventional solar PV cells is the use of low–bandgap materials to exploit low– energy, long–wavelength IR photons. Yang’s group used GaSb cells, which have a bandgap of ~0.8eV and absorb wavelengths up to 1.8μm. While silicon cells are much less expensive,
K. Qiu of Canada’s CANMET Energy Technology Centre said that most thermal radiation falls short of their bandgap (silicon
Eg = 1.12eV). Qiu’s group used a selective Yb2O3 radiator to shift the spectrum of their natural gas burner to the desired range [2].
Control of excess heat is extremely important for successful TPV cells, which encounter far higher temperatures and energy densities than solar PV cells yet only absorb and convert a small fraction of that energy. Most designs reflect excess heat back to the combustion chamber, where it serves to preheat the air–gas mixture and increase the combustion efficiency.
As J. van der Heide and coworkers at IMEC explained, TPV cells generate a higher current density than conventional solar cells; it is important to minimize the resistance of the front contact structure. For the back surface, a highly reflective contact is desirable. Without some form of optical confinement, many long wavelength photons would simply pass through the thin–film cells without being absorbed. The IMEC group, working with germanium TPV cells (Eg = 0.66eV), used an amorphous silicon/SiO2/aluminum stack to form this reflector, realizing the contact structures by laser firing. For the front surface, they diffused palladium from a thin coating through the silicon passivation to form the contact, with a thick silver layer on top to reduce series resistance [3].
In many applications, TPV cells will compete with batteries or fuel cells, not conventional generators or solar PV. Energy density will be as important as cost or net efficiency. Contact resistance, optical confinement, and cell efficiency contribute to energy density as well. TPV and solar PV designs are likely to learn from each other, enriching the range of PV applications in the process. ???K.D.
References:
- W. M. Yang et al., “Experimental study of micro–thermophotovoltaic systems with different combustor configurations,”Energy Conv. and Mgmt., Vol. 48, 2007, pp. 1238–1244.
- K. Qiu and A.C.S. Hayden, “Development of a silicon concentrator solar cell based TPV power system,” Energy Conv. and Mgmt., Vol. 47, 2006,
pp. 365–376. - J. van der Heide et al., “Optimisation and characterisation of contact structures for germanium thermophotovoltaic cells,” presented at 22nd EU PVSEC, 3–7 September, 2007, Milan.